cherryteastain

A related viewpoint is that overparametrization is good because the model is stranded when the Hessian has all positive/zero eigenvalues. If we treat the probability that a particular Hessian eigenvalue turns positive as a Bernoulli process, the chance of all eigenvalues going positive/zero exponentially decreases as the parameter count increases

[1] https://arxiv.org/abs/1406.2572

> shameful to work for

As bad as Meta products are for society, I'd say Palantir is far more shameful to work for.

> If you're building out a brand new system, why not make use of the computing device with input/output capabilities

There is no sovereign EU mobile OS. Adding a hard dependency to Android/iOS is removing one US hard dependency in your payment stack to add another. So, physical cards are a must until there is a European mobile OS with widespread adoption (i.e. never).

Unfortunately, the EU approach here has been not only adding this iOS/Android hard dependency, but also locking it down with crap like device attestation to make sure that it is impossible to use their "sovereign" systems without a US corporation (Google/Apple) certified device. They are actively hostile to alternatives like Lineage or Graphene for instance.

Agree re. prediction markets and predatory marketing but disagree so hard with this

> The private interest is genuine, a global market's appetite for a frictionless way to hold dollars, captured by the saver who holds the token and the issuer who books the reserves. The cost is paid by everyone outside that transaction. What looks rational for the individual Nigerian saver is corrosive for Nigeria.

The way this is framed by the author is something like "poor $COUNTRY central bank has its citizens best interests at heart but evil stablecoins are tying the poor central bank's hands". The reality could not be farther from truth. In countries mentioned in the article like Argentina, Turkey or Nigeria the governments are incredibly corrupt and they use monetary policy and capital controls to make loads of cheap financing available to the ultra rich while inflating their debts away. The net effect is that in these countries the combination of inflation and currency debasement is used as a direct wealth siphon from the middle/upper middle class to the ultra rich (the poor have no savings and therefore are less affected). As a result the middle and upper middle classes of these countries entirely evaporated in the last 10-15 years.

Stablecoins are not the issue here, the governments are.

Crackpot really has connotations like "flat earther" and "aliens built the pyramids". It's one thing to say "I believe GA proponents' claims regarding the usefulness of the geometric product are overstated". It's another to say "GA proponents are crackpots".

Not a fan of the article. It resorts to ad hominem attacks like

> GA had gotten a bad reputation because of its tendency to attract bad mathematicians and full-on crackpots. Hestenes honestly sounds like one a lot of the time, and I’m not really sure whether he is or isn’t. It makes sense, really.

> GA ended up appealing to a lot of fringes: people who only had undergraduate degrees, people who had dropped out of PhDs, people with PhDs from unrigorous programs, people who had been good at math but were perhaps going a bit senile, random passerbies from engineering or computer programming, run-of-the-mill circle-squarers, people who had a bone to pick with establishment mathematics and felt like all dissenting views were being unfairly suppressed

> It didn’t help that a lot of the texts by the actually-competent GA people, like the Cambridge group, tended to say things that sounded and still sound kind of crackpotty as well.

After reading the article, the main "case against geometric algebra" I could find in there was that the author does not like the people using/doing research in geometric algebra, such as the ostensibly failed academics from a Cambridge research group [1] which the article links to.

I was expecting in the "An Actual Case Against GA" section that the author would demonstrate something like "Geometric Product actually does not work if you apply it to xyz domain". Rather, the section just ended up being mostly about the type of bikeshedding you see about naming of variables in programming.

There is I guess merit to the core "there is no good general interpretation or usage for the geometric product or mixed-grade multivectors" thesis of the article but calling other academics crackpots really subtracts from that message.

[1] https://corde.phy.cam.ac.uk/

You can statically link boost

Mullvad already does all of the things you mentioned. Cryptocurrency payments with no KYC, and many anti-censorship [1] methods. Moreover, their mobile and desktop apps apply heuristics to find combinations of anti-censorship measures that work automatically. It's very user friendly (all the user needs to do is press on connect). So, the blocking attempts are still pointless because large, prominent providers already make all these bypass methods very user friendly.

[1] https://mullvad.net/en/help/connecting-to-mullvad-vpn-from-r...

> the bursts can be correlated

Hawkes processes are what other fields use to model this

You may want to look into coroutines so that the data can be fetched asynchronously.

China managed it by keeping US tech out despite, initially, not having alternatives to Google et al.

In winner takes all industries you MUST be protectionist and develop domestic alternatives.

Very useful feature for trading systems where the exchange might tell you your order got filled but then send a trade correction or bust message

This logic works only if distilling Claude is the only way to create another SOTA LLM, which is not the case.

I guess this is an anti-distillation move?

Do they still charge you $50/MTok?

If so, it sounds like a scam. If not, distillers will know which model they are getting by just looking at their API usage.

> reversing it does not seem to benefit the people who were harmed by it

Second and third laws of thermodynamics - our universe has no (macroscopic scale) reversible processes and every irreversible process causes losses

In C++ you'd just do

    template <typename T>
    void DoSomething(T const& data);

or if T* is supposed to point to a tightly packed buffer

    template <typename T>
    void DoSomething(std::span<T> data);

as the author pointed out. I don't see how that is ugly or more complicated than the original void* approach.

There is no need to pass the size of T or length of the span, former is just a sizeof(T) away and latter is a data.size(); away.

In fact, a lot of codebases would outright ban the uint8_t* and reinterpret_cast trick the author is complaining about via clang-tidy rules.

Pix is really good but I don't think it beats Solana or Ethereum L2s.

Most recent indicator of peak Pix transaction volumes I could find [1] was 227M/day (=2700 TPS). You can see yourself that Solana does 130-140M/day consistently. Pix fees you quote are still triple Solana's.

Not to mention there is the entire decentralization aspect, which means the government does not control your money as with other blockchains.

[1] https://agenciabrasil.ebc.com.br/en/economia/noticia/2024-09...

[2] https://blockworks.com/analytics/solana/solana-onchain-activ...

Solana mainnet was shown to have capability to handle 100k transactions per second [1]. Granted, it was a synthetic benchmark of noops, but it shows the capability is there. For reference, Visa claim 83k TPS [2] for their system. It also has the advantage of being vastly cheaper than tradfi payment networks at ca. $0.0005 per transaction [3] irrespective of transaction amount.

[1] https://www.coindesk.com/markets/2025/08/18/solana-briefly-h...

[2] https://corporate.visa.com/en/sites/visa-perspectives/securi...

[3] https://solana.com/learn/understanding-solana-transaction-fe...

> All payment systems are centralized

Except for blockchain based ones

Treat this as a "what do I need to do to make fluid flow that looks okayish in graphics" post rather than "how do I implement physics accurate CFD for industrial/scientific purposes" post.

Lots of points in there like

> "Air is an example of a compressible fluid; you squish it and it gets smaller. Water is an example of an incompressible fluid; you squish it and it pushes back, and doesn't get any smaller" (this only really depends on the Mach number, Ma>~0.3 and you are in compressible territory for any fluid. Incompressibility usually means we assume the divergence of velocity of zero)

> "Incompressible fluids are simpler to simulate because their density and pressure is always constant." (This is only true if you choose to adopt a grad P = 0 approximation)

are incorrect from a physics perspective.

If you look at what real incompressible Navier-Stokes solvers do [1], it's mathematically totally different from what this post shows. In fact, the part that this post omits (handle the pressure gradient term by first approximate time stepping the velocity term by ignoring the pressure term and then correct by solving a Poisson equation for the pressure residual, and then correct the velocity) is the most expensive step in incompressible solvers by far.

[1] https://en.wikipedia.org/wiki/Projection_method_(fluid_dynam...

From what I understand the author was already using Telegram. To avoid this issue, you can install and set up Telegram Desktop before losing your phone and receive the activation code there when setting up your new phone.

It's more useful to view the whole situation as EU politicians prioritizing to have their pockets filled with lobbyist money, rather than the EU as a political entity deciding this per se.

Surely expelling more effective from the school's perspective.

In practice if you use no US part you will get sanctioned as a 'national security risk' or whatever anyway. Even if your product uses no US parts the customers who still want to interact with the US financial system (by e.g. having a USD account) will still be unable to purchase it.

Seems to solve a problem very similar to Conan or vcpkg but without its own package archive or build scripts. In general, unlike Cargo/Rust, many C/C++ projects dynamically link libraries and often require complex Makefile/shell script etc magic to discover and optionally build their dependencies.

How does craft handle these 'diamond' patterns where 2 dependencies may depend on versions of the same library as transitive dependencies (either for static or dynamic linking or as header-only includes) without custom build scripts like the Conan approach?

Curious that they listed the GPU as "RDNA5 AT0 XL". RDNA5 is not out yet, but that naming seems to align with the naming in this leak [1]. It's explicitly labelled as a "desktop/gaming" model with 154 CUs and 36GB of VRAM.

[1] https://www.club386.com/wp-content/uploads/2025/07/rdna-5-ml...

We can see how monumentally important to preserving the right to privacy the work to stop Chat Control has been by the frothing anger we see displayed here. The companies which complain to Uncle Sam every time they're fined by the EU for getting caught red handed smothering competition now ask the EU for more regulation.

Ms. von der Leyen will need to find a way to make things up for Google et al considering she has been pushing this regulation at their behest with fervour. That will probably take the form of even further prostration of Europe to US big tech.

What's the 'new normality' in the fifth stage? Do you think you'll start to believe it actually works 100%? Or that you won't change your assessment that it works only sometimes, but maybe pulling the lever on the slot machine repeatedly is better/more efficient than doing it yourself?

Presumably they meant WW2 Germany.